Automatic Learning for Dynamic Markov Fields with Application to Epidemiology

Following an outline of dynamic Markov fields, we briefly describe some spatial models for contagious diseases and pose a prototype epidemic control problem. The notion of automatic learning is then introduced, and its relevance to epidemic control is described. In essence, once a contagion model is adopted and a domain of controls has been selected, learning can be used to obtain asymptotically optimal performance. (The learning algorithm is a synthesis of simulation and optimization, and is a suitable alternative to response surface methodology, in many applications.) The end product is the same optimal control as would be obtained by a conventional analysis. The point is that our current understanding of dynamic Markov fields does not permit conventional analysis; automatic learning has no computationally competitive alternative. The theory is illustrated by application to a spatial epidemic control problem.